With a new model that matches the performance of ChatGPT o1, DeepSeek managed to turn restrictions into innovation.